specific method · filed under model architecture
Output-independent MoE gradient reformulation
A gradient transformation that depends on the intermediate activation and upstream gradient, eliminating backward dependence on the output.
- source
- 1
- model
- 1
- lab adopt it
- 1
- strongest
- core
How sources treat it
One count per evidence span, weakest treatment to strongest.
core 1
Documented in
Evidence
1 span quoted from the sources, strongest treatment first.
we rewrite this gradient through a mathematical transformation into a form that depends only on the intermediate activation act_output and the upstream gradient doutput, eliminating the backward dependency on output
coresoftware implementationin Kimi K3Moonshot AI
Filed alongside
Other methods under model architecture :: channel mixer :: mixture of experts.
Mixture of ExpertsSparse expert activationDeepSeekMoEGated DeltaNet MoEMixture-of-Experts layersMoE with routed and shared expertsAsymmetric input/output activation splitDispatch recomputationFrequency-based partitioning of N-gram embedding slotsHybrid Mixture of ExpertsMegaMoEMoE with 128 routed experts and a shared expertMoE with 256 experts and top-8 routingMoE with 384 routed experts and a shared expertRouted expert output modulationSigmoid-based MoE router with auxiliary-loss-free load balancing